Research claims that models can transfer complex traits such as new skills, agent hacking behavior, and backdoors.
AI summary
Researcher Owain Evans introduced his new paper, building on previous work on subliminal learning, demonstrating that more complex traits can be transferred between models, including new skills, agent hacking behavior, and model backdoors. The authors noted that during the transfer of backdoors, neither triggers nor corresponding behaviors were present in the training data. Previous research showed that models could transfer a preference for owls through number sequences.
Full text · Original
New paper.
Previous: Models transfer a love of owls through sequences of numbers (Subliminal Learning).
New: We show models can transfer more complex traits: novel skills, agentic hacking, backdoors. Here, a backdoor transfers with neither trigger nor behavior in the data!